Virtual assistant architecture for natural language understanding in a customer service system

Abstract

A virtual assistant system for communicating with customers uses human intelligence to correct any errors in the system AI, while collecting data for machine learning and future improvements for more automation. The system may use a modular design, with separate components for carrying out different system functions and sub-functions, and with frameworks for selecting the component best able to respond to a given customer conversation.

What is claimed is: 
     
       1. A computer-implemented method comprising:
 training dialog components to interpret user utterances based on training data including utterance data obtained from at least one of human intelligence (HI) or artificial intelligence (AI), wherein training the dialog components comprises: 
 for a first task corresponding to a dialog component: 
 for prior stored conversations, comparing results obtained by a first task model corresponding to the first task with stored results obtained from human agents who previously handled the conversations, and 
 responsive to the comparison indicating sufficient confidence in the first task model, adding the dialog component to a set of dialog components eligible for runtime use; 
 establishing a conversation between a user and a virtual assistant such that the virtual assistant manages the conversation, the conversation having corresponding dialog data including a state of the conversation; 
 identifying a task within the conversation, the task being associated with one or more of the trained dialog components and having a corresponding task model; 
 determining, based on a confidence score from application of the task model for the task being below a threshold, that the virtual assistant is unable to complete the task; 
 responsive to determining that the virtual assistant is unable to complete the task, assigning the task to a human agent for completion in an asynchronous manner, without the virtual assistant transferring control of the conversation to the human agent, such that the virtual assistant interprets a next task of the user in the conversation while the human agent is interpreting the task; 
 the virtual assistant receiving a result of the task from the human agent; and 
 the virtual assistant using the result of the task to respond to the user in the conversation. 
 
     
     
       2. The computer-implemented method of  claim 1 , wherein the virtual assistant continues with the conversation while waiting to receive a result of completion of the task by the human intelligence. 
     
     
       3. The computer-implemented method of  claim 2 , wherein the human intelligence is a human agent, the computer-implemented method further comprising:
 the virtual assistant joining the human agent into the conversation, such that the user directs conversation to the human agent; 
 the virtual assistant interpreting an utterance of the user as a second task to be performed; and 
 the virtual assistant assigning the second task to human intelligence or artificial intelligence for completion. 
 
     
     
       4. The computer-implemented method of  claim 1 , wherein the virtual assistant manages the conversation while waiting to receive a result of completion of the task by the human intelligence. 
     
     
       5. The computer-implemented method of  claim 4 , wherein the human intelligence is a human agent, the computer-implemented method further comprising:
 the virtual assistant joining the human agent into the conversation, such that the user directs conversation to the human agent; 
 the virtual assistant interpreting an utterance of the user as a second task to be performed; and 
 the virtual assistant assigning the second task to human intelligence or artificial intelligence for completion. 
 
     
     
       6. The computer-implemented method of  claim 1 , wherein the human intelligence is a human agent, the computer-implemented method further comprising:
 the virtual assistant joining the human agent into the conversation, such that the user directs conversation to the human agent; 
 the virtual assistant interpreting an utterance of the user in the conversation; and 
 the virtual assistant assisting the human agent by providing data to display within an agent desktop user interface of the human agent, based on the interpreting. 
 
     
     
       7. The computer-implemented method of  claim 1 , wherein each trained dialog component has a corresponding task model. 
     
     
       8. The computer-implemented method of  claim 1 , wherein assigning the task to a human agent is additionally responsive to a current queue length for agents. 
     
     
       9. The computer-implemented method of  claim 1 , wherein assigning the task to a human agent comprises matching dialog requirements with characteristics of candidate human agents. 
     
     
       10. The computer-implemented method of  claim 1 , further providing comprising providing the human agent with a graphical user interface allowing the human agent to specify an intent for an utterance in the conversation. 
     
     
       11. A computer system comprising:
 a computer processor; and 
 a computer-readable storage medium storing instructions that when executed by a processor perform actions comprising: 
 training dialog components to interpret user utterances based on training data including utterance data obtained from at least one of human intelligence (HI) or artificial intelligence (AI), wherein training the dialog components comprises: 
 for a first task corresponding to a dialog component: 
 for prior stored conversations, comparing results obtained by a first task model corresponding to the first task with stored results obtained from human agents who previously handled the conversations, and 
 responsive to the comparison indicating sufficient confidence in the first task model, adding the dialog component to a set of dialog components eligible for runtime use; 
 establishing a conversation between a user and a virtual assistant such that the virtual assistant manages the conversation, the conversation having corresponding dialog data including a state of the conversation; 
 identifying a task within the conversation, the task being associated with one or more of the trained dialog components and having a corresponding task model; 
 determining, based on a confidence score from application of the task model for the task being below a threshold, that the virtual assistant is unable to complete the task; 
 responsive to determining that the virtual assistant is unable to complete the task, assigning the task to a human agent for completion in an asynchronous manner, without the virtual assistant transferring control of the conversation to the human agent, such that the virtual assistant interprets a next task of the user in the conversation while the human agent is interpreting the task; 
 the virtual assistant receiving a result of the task from the human agent; and 
 the virtual assistant using the result of the task to respond to the user in the conversation. 
 
     
     
       12. The computer system of  claim 11 , wherein the virtual assistant continues with the conversation while waiting to receive a result of completion of the task by the human intelligence. 
     
     
       13. The computer system of  claim 12 , wherein the human intelligence is a human agent, the actions further comprising:
 the virtual assistant joining the human agent into the conversation, such that the user directs conversation to the human agent; 
 the virtual assistant interpreting an utterance of the user as a second task to be performed; and 
 the virtual assistant assigning the second task to human intelligence or artificial intelligence for completion. 
 
     
     
       14. The computer system of  claim 11 , wherein the virtual assistant manages the conversation while waiting to receive a result of completion of the task by the human intelligence. 
     
     
       15. The computer system of  claim 11 , wherein the human intelligence is a human agent, the actions further comprising:
 the virtual assistant joining the human agent into the conversation, such that the user directs conversation to the human agent; 
 the virtual assistant interpreting an utterance of the user in the conversation; and 
 the virtual assistant assisting the human agent by providing data to display within an agent desktop user interface of the human agent, based on the interpreting. 
 
     
     
       16. The computer system of  claim 11 , wherein each trained dialog component has a corresponding task model. 
     
     
       17. The computer system of  claim 11 , wherein assigning the task to a human agent is additionally responsive to a current queue length for agents. 
     
     
       18. The computer system of  claim 11 , wherein assigning the task to a human agent comprises matching dialog requirements with characteristics of candidate human agents. 
     
     
       19. The computer system of  claim 11 , further providing comprising providing the human agent with a graphical user interface allowing the human agent to specify an intent for an utterance in the conversation. 
     
     
       20. A non-transitory computer-readable storage medium storing instructions that when executed by a processor perform actions comprising:
 training dialog components to interpret user utterances based on training data including utterance data obtained from at least one of human intelligence (HI) or artificial intelligence (AI), wherein training the dialog components comprises: 
 for a first task corresponding to a dialog component: 
 for prior stored conversations, comparing results obtained by a first task model corresponding to the first task with stored results obtained from human agents who previously handled the conversations, and 
 responsive to the comparison indicating sufficient confidence in the first task model, adding the dialog component to a set of dialog components eligible for runtime use; 
 establishing a conversation between a user and a virtual assistant such that the virtual assistant manages the conversation, the conversation having corresponding dialog data including a state of the conversation; 
 identifying a task within the conversation, the task being associated with one or more of the trained dialog components and having a corresponding task model; 
 determining, based on a confidence score from application of the task model for the task being below a threshold, that the virtual assistant is unable to complete the task; 
 responsive to determining that the virtual assistant is unable to complete the task, assigning the task to a human agent for completion in an asynchronous manner, without the virtual assistant transferring control of the conversation to the human agent, such that the virtual assistant interprets a next task of the user in the conversation while the human agent is interpreting the task; 
 the virtual assistant receiving a result of the task from the human agent; and 
 the virtual assistant using the result of the task to respond to the user in the conversation.